Low Distortion Noise Cancellers – Revival of a Classical Technique

This chapter presents low-distortion noise cancellers with their applications to communications and speech recognition. This classical technique, originally proposed by Widrow et al. in mid 70’s, is first reviewed from a view point of output-signal distortion to show that interference and crosstalk are the primary reasons. As a solution to the interference problem, a paired filter (PF) structure introduces an auxiliary adaptive filter for estimating a signal-to-noise ratio (SNR) that is used to control the coefficient-adaptation stepsize in the main adaptive filter. A small stepsize for high SNRs, when the desired signal seriously interferes the misadjustment, provides steady and accurate change of coefficients, leading to low-distortion. This PF structure is extended to more general cases in which crosstalk from the desired-signal source to the auxiliary microphone is not negligible. A cross-coupled paired filter (CCPF) structure and its generalized version are solutions that employ another set of paired filters. The generalized CCPF (GCCPF) is applied to speech recognition in a human-robot communication scenario where improvement in distortion is successfully demonstrated by evaluations in the real environment. This robot had been demonstrated for six months at 2005 World Exposition in Aichi, Japan.